TBNAI
HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 04/10/2026, 20:00:00 UTC
Briefing created: 20/10/2026, 20:55:50 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 13:18:24 UTC.
Briefings
This study introduces HLA-WM, a training-free framework for long-horizon video world models that combines geometry-guided retrieval with recurrent linear attention. On the 60-second SANA-WM-Bench, HLA-WM improved all six revisit-consistency and camera-control metrics of the base autoregressive generator, achieving a 0.74 dB PSNR gain and a 28.5% reduction in rotation error. The method reduces historical-state memory by 12 times relative to full KV caching while maintaining inference throughput within 1.6% of baseline performance. Results also generalize to MBench-A across four inference modes over 547 samples.
Key points
- HLA-WM improves all six aggregate revisit-consistency and camera-control metrics on the 60-second SANA-WM-Bench without additional training.
- HLA-WM reduces historical-state memory by 12 times relative to full KV caching while incurring at most a 1.6% reduction in inference throughput.
Context
Video world models require persistent memory to preserve scene consistency over extended rollouts. Softmax attention retains full generation history, while recurrent linear attention compresses history into fixed-size states. However, Gated DeltaNet suffers from severe long-range forgetting where information from distant but relevant scenes is progressively attenuated by subsequent state updates.
Limitations
Results are limited to the 60-second SANA-WM-Bench and MBench-A evaluations. Memory reduction is measured relative to full KV caching, and inference throughput reduction is capped at 1.6%. The study does not report results for other video lengths or specific inference modes beyond the evaluated subsets.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 20/10/2026, 20:55:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidgitkcu2ng6we5c6lmr7hkuuw7x224k64gfefvy6fei5gf7gw5u4